lomahony/pythia-70m-helpful-sft

Pythia-70m supervised finetuned using TRLx library with the helpful subset of Anthropic-hh-rlhf dataset for 1 epoch.

Checkpoints are also uploaded.

Fully reproducible finetuning code is available on GitHub

See Pythia-70m for model details (paper).

hf (pretrained=lomahony/pythia-70m-helpful-sft), gen_kwargs: (None), limit: None, num_fewshot: 0, batch_size: 16

Tasks	Version	Filter	Metric	Value		Stderr
arc_challenge	1	none	acc	0.1715	±	0.0110
		none	acc_norm	0.2082	±	0.0119
arc_easy	1	none	acc	0.3384	±	0.0097
		none	acc_norm	0.3262	±	0.0096
boolq	2	none	acc	0.4239	±	0.0086
hellaswag	1	none	acc	0.2629	±	0.0044
		none	acc_norm	0.2691	±	0.0044
lambada_openai	1	none	perplexity	5937.7964	±	424.7555
		none	acc	0.0328	±	0.0025
openbookqa	1	none	acc	0.1580	±	0.0163
		none	acc_norm	0.2520	±	0.0194
piqa	1	none	acc	0.5593	±	0.0116
		none	acc_norm	0.5392	±	0.0116
sciq	1	none	acc	0.3710	±	0.0153
		none	acc_norm	0.4990	±	0.0158
wikitext	2	none	word_perplexity	550.5954	±	N/A
		none	byte_perplexity	3.2550	±	N/A
		none	bits_per_byte	1.7027	±	N/A
winogrande	1	none	acc	0.4878	±	0.0140

hf (pretrained=lomahony/pythia-70m-helpful-sft), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: 16

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
arc_challenge	1	none	5	acc	0.1869	±	0.0114
		none	5	acc_norm	0.2210	±	0.0121
arc_easy	1	none	5	acc	0.3207	±	0.0096
		none	5	acc_norm	0.3245	±	0.0096
boolq	2	none	5	acc	0.4159	±	0.0086
hellaswag	1	none	5	acc	0.2633	±	0.0044
		none	5	acc_norm	0.2596	±	0.0044
lambada_openai	1	none	5	perplexity	19968.0749	±	1423.3001
		none	5	acc	0.0202	±	0.0020
openbookqa	1	none	5	acc	0.1440	±	0.0157
		none	5	acc_norm	0.2420	±	0.0192
piqa	1	none	5	acc	0.5359	±	0.0116
		none	5	acc_norm	0.5229	±	0.0117
sciq	1	none	5	acc	0.3240	±	0.0148
		none	5	acc_norm	0.4310	±	0.0157
wikitext	2	none	5	word_perplexity	550.5954	±	N/A
		none	5	byte_perplexity	3.2550	±	N/A
		none	5	bits_per_byte	1.7027	±	N/A
winogrande	1	none	5	acc	0.5154	±	0.0140

lomahony
/

pythia-70m-helpful-sft